A step-by-step playbook for an AI agent that answers product questions from your own pages, brochures and price sheets, without guessing.
Picture a procurement engineer on your website at 10 pm on a Thursday. They need to know whether your pump comes in stainless steel, what the lead time is for 40 units, and whether it fits a 2-inch flange. The answers are somewhere on your site. One is in a PDF brochure, one is on a product page, and one is in a price sheet your sales team emails out. They give up after five minutes and open a competitor's tab.
An AI agent that answers product questions from your own content fixes that particular problem. It won't close the deal for you, but it can stop people leaving before they've even found out what you sell.
I build AI chat agents for a living, so I'm not neutral about this. What I can offer is the setup we actually use, including the parts that go wrong.
What the research says about buyers
Gartner surveyed 632 B2B buyers in August and September 2024. It found that 61% prefer an overall rep-free buying experience. In the same survey, 69% reported inconsistencies between the information on a supplier's website and what the seller told them.
To be fair to sales teams, that doesn't mean buyers want no people at all. In the same release, Gartner analyst Alice Walmesley says sellers should offer unique guidance and act as a sounding board for buyers. My reading: buyers want the facts on their own schedule, and a person when the question gets hard. A product Q&A agent is for the first part. It should hand over for the second.
What a product Q&A agent does (and doesn't)
A good one answers questions about your products and services from the material you give it: product pages, brochures, manuals, spec sheets and price lists. It points people to the right page. When it doesn't know, it says so and offers a person.
It doesn't invent a price, a lead time or a compatibility claim that isn't in your content. If you take one rule from this post, take that one.
Step 1: Build the knowledge base first
The agent is only as good as what it can read. In AtomicX, that means a knowledge base, under AI → Knowledge bases.
- Upload your documents. Brochures, manuals and price sheets go in on the Documents tab. TXT, Markdown, PDF, DOCX and HTML files are supported, up to 10 MB each and 20 at a time.
- Add your website. On the Websites tab, add your site and choose a Maximum Depth. Start low and go deeper only if important pages are missing.
- Set a Rescan frequency. Daily, Weekly or Monthly keeps answers in line with your site. The default is Manual only, which is easy to forget.
- Check what the AI can actually read. Click View on a document, or Pages on a website, to see the extracted text. If a spec table didn't come through, the agent can't quote it.
Two small settings are worth knowing. Tick Available for download on a document, and visitors who get an answer from it see a download link below the reply. And product pages with Open Graph image and description tags show up as tidy link cards under relevant answers.
Step 2: Create the agent
You can write an agent from scratch under AI → AI agents, or start from the Agent Gallery. In the gallery, click Add to my agents and a copy lands in your own list as a Draft. Treat it as a starting point. Add your company name, products and rules before you publish.
A few settings matter for product questions:
- AI model. The Regular model is the right choice for most sales and support chats. The Premium model handles complex, multi-step questions better, but uses more credits and replies more slowly.
- Temperature. Lower is safer and more factual. For regulated industries, AtomicX recommends 0.4 or lower. I'd stay low for anything involving specs or prices.
Step 3: Write the instructions
The Agent Instructions (System Prompt) box is where most of the quality comes from. Specific rules work better than vague ones. Here's an example I'd start from for a hypothetical industrial equipment maker. Adapt it to your business.
- You are the product assistant for Northfield Pumps. You help engineers and buyers find the right product and the facts about it.
- Answer only from the knowledge base. If the answer isn't there, say you don't know and offer to connect the visitor with our sales engineers.
- Never guess prices, lead times, certifications or compatibility. Quote them exactly as written, and say where they come from.
- If a price depends on quantity or configuration, say so and offer a quote from the team.
- Keep answers short: two to four sentences, then offer more detail.
- Ask one clarifying question when a request is ambiguous, for example which model or which size.
- Don't compare us with named competitors.
- Be professional and plain-spoken. No sales pressure.
If you're stuck, Improve with AI will rewrite your draft with a clearer role, tone and rules. It replaces what's in the box, so read it before you save.
Step 4: Put it in a workflow
Visitors don't pick an agent. You attach it to an AI Chat node in a workflow, under Chat → Workflows. In the node's AI Configuration, choose the AI Agent and the Knowledge Base. If you leave the knowledge base empty, the panel warns you that the agent will answer without your content. Take that warning seriously.
Then plan the exits. Add a Handoff node for questions that need a person, and turn on Enable exit detection so the AI can move the visitor there when it fits. If you want meetings, add a Book Appointment node too. Only published agents appear in the node, so remember to click Publish.
Step 5: Test it like a sceptical buyer
Go to Insights → Chat test and ask the questions your buyers actually ask. Then ask the awkward ones:
- A question that needs facts from two different documents.
- A question your content doesn't answer. The right reply is "I don't know", not a guess.
- A price question for an unusual quantity.
- The same questions in another language. The AI replies in the visitor's language, so check names, numbers and units.
Test chats are real conversations, so they show up in your conversation list. Ask a colleague who didn't build it to try too.
How to measure it
Start with Insights → Analytics. The Completion Rate and the Handoffs column in Detailed Statistics tell you a lot. Many handoffs on the same topic usually mean the knowledge base is missing something.
Then read conversations. In Chat → Conversations, click Generate analysis for scores on Relevance, Helpfulness, Tone, Resolution and Error handling, with suggestions. On any AI reply, Show AI Details shows the agent version and the knowledge base context it used. That's how you tell a bad document from a bad instruction.
Visitor thumbs up and down appear under Feedback. And AI → Improvements turns negative ratings and failures into proposed changes your team reviews and tests before anything goes live.
Pitfalls I see most often
- Old price sheets. If last year's PDF is in the knowledge base, the agent will quote it. Delete superseded files.
- Forgetting to publish. Chats keep using the last published version of an agent until you click Publish again.
- Syncing over your changes. When a gallery agent shows Update Available, Sync overwrites your custom instructions. Copy them first.
- Scanning too deep. A high depth pulls in job ads, old news and legal pages. They dilute the answers.
- No exit. An agent with no handoff path frustrates the exact buyers you most want to talk to.
The bigger pattern
Gartner's 69% figure is the one I'd keep in mind. Buyers notice when your website, your brochure and your salesperson tell three different stories. An agent that answers from one set of documents won't fix messy content. But it will show you, conversation by conversation, where the mess is.
Start small: one knowledge base, one agent, one workflow, and the 20 questions your sales team hears most. Want to see it on your own content? Book a 20-minute demo, or read the AI agents guide first.